Introduction

Row

Valuebox

100

Valuebox

60

Valuebox

80

Valuebox

75

Row

Row

Introduction

When hosting a party or event, having uninterrupted music with smooth transitions is crucial. The Spotify Automix fuctionality can be a good solution, although not all playlists are suitable for this feature. Spotify offers pre-programmed playlists that include the Automix functionality, but users must configure their settings correctly to use it. To learn how to set up Automix functionality properly, refer to this link.

Despite its benefits, Spotify’s Automix functionality has some limitations compared to a regular DJ. Firstly, it lacks the complete experience of a DJ. Secondly, users cannot easily add songs to pre-programmed playlists. Lastly, there has been limited research on the techniques used by the Automix functionality in music studies. This raises the question of whether the techniques used by the Automix functionality are similar to those used by regular DJs or not.

This portfolio investigates the relationship between Spotify’s Automix functionality and the mix components used in the DJ industry for smooth song transitions.The Automix playlists used for this portfolio are HouseWerk, TechnoBunker, Mint and Dance Rising. The music contained in the playlist is best classified under House music. In subcategory the Housewerk playlist contains techhouse, TechnoBunker techno, DanceRising and Mint dance pop. The interest in this comparison came from being a DJ myself. In the DJ world, mainly 2 components are used that are important for a transition between two songs. These are the tempo of the songs and the key that the mixing songs are in. It will be investigated whether these components are also used by the automix functionality. This will focus on the order of the songs in the playlist. The shuffle functionality will be left out. It may also turn out that the functionality is completely random.

Picture of me behind the decks

Tempo

How is the tempo distributed in the playlists?

The first visualization shows the distribution over tempo of the songs within the different playlists. In the visualization below, the x axis shows the different playlists. On the y axis, the tempo is visible. If we look closely at the visualization we see that the playlist Technobunker has a high median compared to the other playlists. There are also the fewest outliers here compared to the other playlists. The spread of Housewerk is the smallest. This means that the tempo in this playlist is mostly close to the median. This is a positive first analysis on the research question.
The outliers negatively affect overmixing. At least on a basic level of mixing. From the DJ industry, songs with a maximum tempo difference of 6 bpm can be mixed. A so-called Master Tempo function on the turntable ensures that the key does not change at that moment. See a helpful explainer video here. When the tempo difference is greater than 6 bpm, the listener will clearly hear the difference in tempo. A possible solution to compensate for the large bpm difference is to use tempo harmonics and subharmonics. For example, a 120 BPM song can be mixed with the 60 BPM subharmonic. In the following visualization it is examined whether this is used.

How does the tempo behave throughout the playlist?

The second visualization shows the progression of tempo over the playlist. On the x axis here the songs are numerically represented in playlist order. On the y axis, the tempo is shown. The figures are analyzed by playlist. The second visualization complements the first visualization well. Both figures clearly show that Dance Rising and mint have the most outliers. Looking at the order, the technique of subharmonics (as discussed on the previous page) is probably used once. In the dance rising playlist this can be seen in around track 45 and 46. The tempo difference between track 45 (80 bpm) and track 46 (160 bpm). Unfortunately, in cases of the large tempo difference, this technique is not used more often. We can therefore already draw a first conclusion that in these cases the transitions in terms of tempo are random.
The best build in tempo can be seen in the technobunker playlist. In this visualization, the pace is clearly building and decreasing. There are few peaks in the visualization, which means that techniques such as subharmonics are not applicable. This is positive for the transitions

An outlier that makes the transition complicated even for a regular DJ.

In the DJ world, the easiest method to mix two songs is to keep the tempo the same throughout the song. When this is not the case, it requires an advanced technique to mix the songs together. The tempo analysis features in the turntable are now no longer accurate because this currently gives an average value of the tempo over time rather than the tempo per section in the song. The DJ must be familiar with this change in tempo in order to mix the song correctly. An example of an outlier in the corpus is the song Affraid to Feel by LF SYSTEM. The song was released on May 2, 2022. The song uses the sample from the song I Can’t Stop (Turning You On) by SILK. This song has a tempo of 85 bpm. So for LF SYSTEM it was a challenge to incorporate the sample into a disco track with a tempo of 130 BPM. This is evident from the tempogram. The song starts slowly with a tempo of 100 BPM. After 20 seconds, the tempo increases to 130 BPM. This is when song switches from the intro to the chorus. What is not immediately clear is that after the chorus, the tempo drops slightly. After repeated listens, this became audible. From the tempogram it can be seen that after the first chorus, the chorus returns 2 more times. The last thing that stands out is that after about 150 seconds the tempo is classified as rapidly increasing. In the recording it can be heard that the kick drum changes from every four-quarter measure to every 8th measure. This makes the tempo seem to double but is not actually the case. It is the lead-up to the chorus.

Key

Energy and Valence

Column

Can valence, energy, loudness and mode give an unexpected addition?

Chromograms

Row

First song Housewerk

In this section, we chose to compare two songs from the playlist Housewerk. The songs chosen are listed in order in the playlist. This is to see if it becomes clear if certain chords are used in the songs that go well together. This makes the transition between the songs sound better. Starting with the first track, the song No Diggity produced by Häwk and Beyge was chosen. The song is a cover of of the original song written by Blackstreet, Dr. Dre and Queen Pan that was released in 1996. The song is played primarily in C# and F#. Chebyshev was chosen as the normalization. This is because with the euclidean and manhattan it was not clearly visible which notes were used.The second song analyzed is the song Ritmo by Raffa FL. It is a cover of the song La Colegiala by The Boy Next Door, Fresh Coast and Jody Bernal. From the chromogram, we notice that the song is played in F, G and Gm. Euclidean was chosen as the normalization. This is because with this method the magnitude (in yellow) was the clearest. With confirmation from the spotify API, it is clear that these two songs can be well mixed together.

Second song Housewerk

In this section, we chose to compare two songs from the playlist Housewerk. The songs chosen are listed in order in the playlist. This is to see if it becomes clear if certain chords are used in the songs that go well together. This makes the transition between the songs sound better. Starting with the first track, the song No Diggity produced by Häwk and Beyge was chosen. The song is a cover of of the original song written by Blackstreet, Dr. Dre and Queen Pan that was released in 1996. The song is played primarily in C# and F#. Chebyshev was chosen as the normalization. This is because with the euclidean and manhattan it was not clearly visible which notes were used.The second song analyzed is the song Ritmo by Raffa FL. It is a cover of the song La Colegiala by The Boy Next Door, Fresh Coast and Jody Bernal. From the chromogram, we notice that the song is played in F, G and Gm. Euclidean was chosen as the normalization. This is because with this method the magnitude (in yellow) was the clearest. With confirmation from the spotify API, it is clear that these two songs can be well mixed together.

Similarity Matrix

La Fuente - I Want You, An interesting homogeneity pattern

Above are the similairty matrices of the chroma features (left) and timbre (right). The song chosen is La Fuente - I Want You. The song is in key 8, the tempo is 125 bpm and energy is 0.8. We analyze the figure using 3 criteria. Representation (shows the repeating elements with diagonal sloping lines), novelty (bright yellow lines when a part is not similar and dark lines when it is) and lastly homogeneity (A so-called chessboard that when it is dark blue shows that everything is similar and when it is more yellow more difference). Starting with the representation, it is difficult to find the diagonal lines. They can be seen very slightly at the end of the song Looking at the novelty, it is clear that the song has 2 refrains. These are the intersections of the yellow lines around 40 and 190 seconds. At that point the song changes tremendously. Looking last at the homogeneity, the chessboards are mainly found between 50 and 100 seconds.

Tom Santa - Rainfall, Checkerboards everywhere

Looking at the second song, we look at the analysis of the song Tom Santa - Rainfall. This is a cover of the song Shackles released by Mary Mary. The song is in key 5, the tempo is 128 bpm and the energy is almost 0.9. Like the song by La Fuente, I analyze this song using the 3 characteristics of representation, novelty and homogeneity. Starting with representation, we see the diagonal dark lines a lot in the song. During the beginning of the song until 30 seconds it is clearly visible. Looking at the novelty, a large intersection can be found around 90 seconds. You can clearly hear the transition in the song. Lastly looking at the homogeneity, the chess boards are also well found in this song. Especially some smaller chessboards between 25 and 90 seconds. Because these are dark blue chessboards, the song changes a lot in this time interval

Clustering

Which features are important for improving mixing?

Using random forest, an attempt was made to obtain the most important features from the playlists. Due to the size of the playlists, the decision was made to focus only on the first 20 songs of each playlist. In order to get an overall view, the choice was made to use as many features as possible. The visualization shows that tempo is most important. This meets the expectations. What is interesting is that the 6th coefficient of the timbre vector is almost as important as tempo. It is difficult to determine what the 6th dimension of the timbre vector means. The documentation on this vector is quite limited. The rest of the features decrease gradually.

Improvements by using clustering

To improve the number of good transitions, I want to show a possible solution using clustering to better determine the order of the playlist. The best features from previous page (tempo and c6) were used to create this clustering. Single linkage was chosen where the smallest distance between clusters is calculated. The choice of this form has two reasons. First, outliers can then be discovered first. These songs can then potentially be used at the beginning of the playlist because the user starts the song at that point. Second, clusters are now created where the distance between songs is smallest. An order of songs in the playlist can now be easily created.

Conclusion and Discussion

Conclusion

The conclusion summarizes the results and answers the research question posed in the introduction.
The introduction introduced the research question: “Does the spotify automix functionality use the same techniques as a regular DJ?”. In this portfolio, visualizations were used to try to answer this question.
Starting with the visualizations on tempo. From the visualizations, it can be seen that the Dance Rising and mint playlist contain the most outliers in terms of tempo. This negatively effects the overmixing of the songs. Housework and Technobunker are the same in terms of tempo. The songs are in the same tempo range and few outliers can be found there. In addition, there is a single song that makes the transition tremendously difficult as seen in the last visualization in terms of tempo. In the song Affraid to feel by LF SYTEMS, the tempo chart showed that the song consists of different tempos that does not make the transition to another song easier.
In the visuals about key, the first insight was generated into how many songs were properly over-mixed based on key. This number was low for all four playlists. Using one example, a good transition was shown and using another example, a bad transition was shown. Using the circle of fifth’s, the difference between a good and a bad transition became easily visible.
From the energy and valence figure, it could be seen that the playlists are very energetic and the valence is very scattered. From the loudness, few outliers could be found. The biggest finding found in this visualization is that the technobunker playlist consists primarily of songs in major.
The Chromograms showed that by eye it is very difficult to tell which chords were used in songs. Therefore, the least data was obtained from this to answer the research question.
The similarity matrices were interesting. It was clear from the first figure that La Fuente’s song I Want You contained multiple choruses but representation was difficult to see. In contrast, the second visualization showed more insight on novelty. The visualizations showed good insight into repetitions and changes in the song. The visualizations did not have a major impact on answering the research question but showed good insights.
The last visualizations in this portfolio regarding the clustering summarize the experiences made in the portfolio. The portfolio showed that the most information regarding the research question could be obtained from the tempo and key components. As mentioned earlier in this conclusion, the transitions made by Spotify’s Automix functionality are not yet optimal. By means of clustering a method is shown how feature selection and tree’s can be used to adjust the order within the playlist in order to improve the number of transitions.

To summarize, we can answer the research question. The tempo and key features show that Spotify makes little use of these features to overmix the songs. A limited number of songs are overmixed well, but that could also be a coincidence. Track popularity will probably have a heavier consideration in the order in which spotify delivers the playlist.In addition, it has also been found that features such as mode, loudness, energy and valence can tell little about the overmix technique. However, these features have provided more insight into the data. From the clustering visualizations, spotify can make a consideration to improve the transitions and enhance the listener’s listening experience.

Discussion

There are a number of discussion points that need to be discussed.

First, the playlist changed over time. Since these were not their own playlists, Spotify changed the playlist every week. As a result, the research result has changed slightly several times. However, the changes are not worth mentioning because most songs remained in the playlist and a single new song was added.

Secondly, Spotify is all about the popularity of a playlist and less about the automix functionality. As a result, the quality of the transitions will weigh less heavily. Spotify may be able to expand playlists where the transition is more important.

Finally, this research was done on only 4 Spotify playlists. For a higher reliability and a more diverse answer, this research could be repeated on the other playlists that contain this functionality.